Chem. Res. Toxicol. 2010, 23, 1215–1222
1215
Developing Structure-Activity Relationships for the Prediction of Hepatotoxicity Nigel Greene,*,† Lilia Fisk,‡ Russell T. Naven,† Reine R. Note,‡,§ Mukesh L. Patel,‡ and Dennis J. Pelletier| Worldwide Medicinal Chemistry and Drug Safety R&D, Pfizer Global Research and DeVelopment, Pfizer Inc., Groton, Connecticut 06340, and Lhasa Limited, 22-23 Blenheim Terrace, Woodhouse Lane, Leeds LS2 9HD, United Kingdom ReceiVed March 4, 2010
Drug-induced liver injury is a major issue of concern and has led to the withdrawal of a significant number of marketed drugs. An understanding of structure-activity relationships (SARs) of chemicals can make a significant contribution to the identification of potential toxic effects early in the drug development process and aid in avoiding such problems. This process can be supported by the use of existing toxicity data and mechanistic understanding of the biological processes for related compounds. In the published literature, this information is often spread across diverse sources and can be varied and unstructured in quality and content. The current work has explored whether it is feasible to collect and use such data for the development of new SARs for the hepatotoxicity endpoint and expand upon the limited information currently available in this area. Reviews of hepatotoxicity data were used to build a structure-searchable database, which was analyzed to identify chemical classes associated with an adverse effect on the liver. Searches of the published literature were then undertaken to identify additional supporting evidence, and the resulting information was incorporated into the database. This collated information was evaluated and used to determine the scope of the SARs for each class identified. Data for over 1266 chemicals were collected, and SARs for 38 classes were developed. The SARs have been implemented as structural alerts using Derek for Windows (DfW), a knowledge-based expert system, to allow clearly supported and transparent predictions. An evaluation exercise performed using a customized DfW version 10 knowledge base demonstrated an overall concordance of 56% and specificity and sensitivity values of 73% and 46%, respectively. The approach taken demonstrates that SARs for complex endpoints can be derived from the published data for use in the in silico toxicity assessment of new compounds. Introduction Drug-induced liver injury (DILI) is a major issue of concern and has led to the withdrawal of a significant number of marketed drugs (1, 2). Adverse effects can range from hepatic enzyme elevations to liver failure (3, 4) and are often difficult to predict in the preclinical stages. For the pharmaceutical industry, hepatotoxicity discovered late in development or after the launch of the drug, leading to its withdrawal, has huge financial implications (5). It is, therefore, important to try to identify these effects early in the drug development process. The etiology of DILI is complex, with many factors including age, disease state, environment, and genetic susceptibility affecting the outcome (6, 7). While some compounds may cause liver damage by direct action, many will act after undergoing metabolic transformation. The role of the liver as the primary site of metabolism and clearance makes it a major target for drug-related adverse effects. The mechanisms involved in detoxifying and facilitating the removal of xenobiotics from the body can also inadvertently lead to their bioactivation to toxic intermediates (7). * To whom correspondence should be addressed. Tel: 860-715-4921. E-mail:
[email protected]. † Worldwide Medicinal Chemistry, Pfizer Inc. ‡ Lhasa Limited. § Current address: L’Ore´al, Life-Sciences/Safety Research Department, 1 Avenue Euge`ne Schueller, F-93600, Aulnay-sous-Bois Cedex, France. | Drug Safety R&D, Pfizer Inc.
Adverse drug reactions in the liver can be classed as intrinsic or idiosyncratic (4). Intrinsic toxicity accounts for some 80% of adverse drug reactions (7), it is predictable, and the effect can be ascribed to some unintended pharmacological action or to the chemical reactivity of the drug or its metabolites. Intrinsic effects will have a dose-response relationship and may be demonstrated across different species. Patient susceptibility to idiosyncratic toxicity cannot be easily predicted in the same way. The occurrence of idiosyncratic toxicity in patient populations is rare, but the effects on the individual can be severe. There are no simple dose-response relationships and no current animal models that can screen for such effects (3, 8, 9). Idiosyncratic toxicity can be divided further into immunologic and nonimmunologic effects. The key signs of an immunologic effect include fever, rash, eosinophilia, the presence of autoantibodies during an initial sensitization period, and a rapid adverse response on rechallenge with the drug (2, 4). Nonimmunological reactions do not show any of these features and can have a longer latency period before revealing their toxicity (2). The doses at which adverse effects can be detected in animal studies or toxicity studies in human volunteers provide valuable information. Dose-response relationships form the basis of drug safety assessment and the determination of a safe efficacious dose (9). Idiosyncrasy thus poses a challenge to this extrapolation. The rarity of occurrence, absence of accepted animal
10.1021/tx1000865 2010 American Chemical Society Published on Web 06/16/2010
1216
Chem. Res. Toxicol., Vol. 23, No. 7, 2010
models, and lack of knowledge of the susceptible population create a challenge in predicting hepatotoxicity. The prediction of adverse effects on the basis of the chemical structure of a compound offers a valuable screening tool early in the drug development process alongside in vitro methodologies (10). Computer methods are well-suited for the purpose of screening large numbers of structures while reducing development time and costs. While such in silico methods have been successfully developed for endpoints such as mutagenicity and skin sensitization, which are based on well-established mechanisms and for which large and diverse data sets are available in the public domain (11, 12), they have not been widely applied to more complex endpoints such as hepatotoxicity (9). For researchers using the published literature, finding good quality data for sufficient compounds to derive such relationships is always problematical. Endpoints such as acute, chronic, or organ toxicity have been difficult to predict, essentially because of the complex nature of the mechanisms involved and the limited number of publicly available data sets (13, 14). Examples published in this area include the use of clinical trial data to develop a quantitative structure-activity relationship (QSAR) model for the prediction of the no observed effect levels (15) and the prediction of nonidiosyncratic human hepatotoxicity using data compiled from the literature to train an ensemble of decision trees (8). Recently, QSAR and expert system-based analyses of a human adverse effects database for hepatobiliary toxicity have been reported (15-17). Knowledge-based expert systems are used to support decision making by emulating human expert reasoning in areas where the knowledge is incomplete and uncertain. These systems consist of two principal parts: the knowledge base and an inference engine (18). A knowledge base will contain a set of facts and related heuristic rules for the domain in question. Heuristic rules are derived from knowledge that has come from experience and in-depth understanding of the domain. Expert systems are designed not only to give a prediction but also to provide the user with the information used by the inference engine to come to its conclusion (18, 19). Derek for Windows (DfW) is a knowledge-based expert system developed at Lhasa Limited to predict the toxicity of a chemical from its structure (20). The knowledge base is composed of structural alerts, example compounds, and rules, which each contribute to and allow for clearly supported and transparent predictions for a number of endpoints (11, 21). Each alert is based on the investigation of available toxicity data for a class of compounds together with other relevant information, including mechanistic understanding, where this is available. The alerts contained in the system have been developed from the published literature and through collaborative efforts with member organizations of Lhasa. While alerts for endpoints such as carcinogenicity, mutagenicity, and skin sensitization are wellestablished, organ toxicity endpoints including hepatotoxicity were poorly represented in the DfW knowledge base. In version 8 of the DfW knowledge base, there were only two alerts for hepatotoxicity, one focused on aliphatic nitro-containing compounds and the other on pyrroline and pyrole esters. We investigated whether it is possible to use the published literature to develop structure-activity relationships (SARs) for chemicals that have the potential to cause hepatotoxicity. The results of the study were implemented as structural alerts in DfW. The creation of a knowledge base for hepatotoxicity and its subsequent development are described.
Greene et al.
Experimental Procedures Computational. DfW versions 8.0.1, 9.0.0, and 10.0.2 were used in the development of the alerts, and all related analyses were run on standard desktop PCs running Microsoft Windows 2000 or XP. Chemical structures were drawn in ISIS/Draw version 2.3 or 2.4, and the databases were created with Microsoft Access 2000 or 2003 provided with Accord version 5.3 (Accelrys Software, Inc.), a plugin for viewing chemical structures. Data Collation and Database Construction. The reference book “The Adverse Effects of Drugs and Other Chemicals on the Liver” (4) was used as the main source of chemical structures for the construction of a structure-searchable database in Microsoft Access. Compounds were also sourced from the reviews by Ludwig and Axelsen (22) and Lee (23). Data related to the compounds were entered into the database and included the compound structure, name, and CAS number and human and preclinical data for hepatotoxicity, where available. Development of Structural Alerts. Chemical structures were exported as an SD file from the database and processed in DfW version 8.0.1 to identify those compounds that already activated an existing hepatotoxicity alert and also to identify relevant structural classes associated with other toxicity endpoints. Additional candidate classes were identified by analyzing the data set by eye and by using the topogenic fragment feature in ChemTK version 3.0.3 (Sage Informatics LLC). Existing therapeutic classes in the reviews were also taken into consideration. The most promising structural classes, in terms of available biological and mechanistic data, were then chosen for further investigation. For each class, toxicological data for component compounds were sought from the original sources cited in the reviews. Searches in TOXNET (http://toxnet.nlm.nih.gov/) and PubMed (http://www.ncbi.nlm.nih.gov/) were carried out to identify any additional supporting data such as human case reports, animal study data, and mechanistic information as well as to identify more compounds in the same structural class. The collated data were then evaluated to establish if there was sufficient evidence to support the development of SARs. Where sufficient evidence was available, SARs were defined and a structural alert implemented in a custom version of the DfW knowledge base. Any additional supporting example compounds that were found during the investigations were added to the compound database along with their toxicological data. Each alert was then tested and peer-reviewed to assess its validity and coverage for the related class of compounds. Following this first stage, those hepatotoxic compounds for which no structural class had been identified were combined with a data set of compounds that were considered negative for hepatotoxicity. The compounds were then clustered using both a maximum common substructure (MCSS) approach and also by the Pipeline Pilot (Accelrys Software, Inc.) Bayesian classification algorithm, using the FCFP_6 structural fingerprint as the descriptor. To identify additional candidate classes, each cluster, as defined by a shared substructure or fragment, was then evaluated for overrepresentation of hepatotoxic compounds and checked to ensure that there was no significant overlap with alerts already developed. A final filter examined the potential impact of the alert by assessing the number of compounds in the Pfizer compound screening files that contained this substructure. Alert classes that would have a low applicability to drug candidates, that is, had few representatives in the Pfizer compound collection, were subsequently deprioritized. These classes were then investigated using the same methodology as described for the first phase, and new structural alerts were implemented in the custom knowledge base, if sufficient evidence was available. Evaluation. During the course of the project, two customized versions of DfW, which included all of the new hepatotoxicity alerts developed up until that point, were created based on the DfW 9 (software version 9.0.0) and 10 (software version 10.0.2) knowledge bases. These were used to evaluate the performance of the new alerts against a Pfizer-developed set of 626 compounds. This set of compounds has significant overlap and a similar classification
Structure-ActiVity Relationships for Hepatotoxicity
Chem. Res. Toxicol., Vol. 23, No. 7, 2010 1217
Figure 1. Example alerts describing SARs developed for tetracyclines and thiophenes.
Figure 2. Drugs containing a thiophene ring and associated with hepatotoxicity.
schema to a set of 344 compounds described previously in the literature (24), although notably with a larger number of compounds. The compounds were assigned one of four classifications associated with hepatotoxicity: NE (no evidence for hepatotoxicity in any species, 152 total), WE [weak evidence for human hepatotoxicity (
